Confirmed lower bound only. Solid lines are complete exact observations; dashed bridges cross intervals without one. Unknown inventory is never plotted as zero.
Planning margin vs API markets
Live rented capacity · OpenRouter context gate · current SiliconFlow + AIHubMix quotes
Model
Market endpoint
Modeled rental cost
Gn delivery
Margin range
Base margin
Break-even rent
Decision
Weekly open-weight candidate audit
Awaiting OpenRouter and GPU price snapshots.
Prompt-cache discounts by provider and model
Live selected routes · follows the model selector above · missing price ≠ zero discount
Agent-session hit-rate sensitivity
Select a model to compare steady-state cache reads.
Cache-hit share
Effective API input $/M
API savings
Modeled self-host input $/M
Input margin
Route / provider
Model
Input $/M
Cache read $/M
Read discount
Cache write $/M
Write premium
Mode
Steady-state read scenarios only. Cache-write charges, TTL/eviction, and the real traffic-wide hit share are not amortized. Self-host cache cost appears only after a matched warm-prefix benchmark. Missing published cache price is unavailable, not zero.
Benchmark next: exact stock
Available classes still missing a full precision profile
Prioritized from exact inventory: validate the SKU, rent the cheapest confirmed offer, then require two clean benchmark runs. Stock is a lower bound, not a reservation.
External rental benchmarks
Settled reference rates versus confirmed provider floors
GPU
Ornn OCPI
Silicon NeoCloud
Silicon hyperscaler
Raw provider floor
GPU Tracker floor
Tracker listings
Latest index date
Shows externally indexed GPUs by default. Expand only when comparing tracker-only listings. Reference rates remain separate from executable offers.
Partner market references
Licensed benchmarks · delivery windows · attributed broker observations
Source / product
Scope and delivery
Native quote
Quantity reference
Evidence / comparison
GX benchmarks and SF orderbook depth are reference data. SF quantities describe purchasable schedules for one window, with cross-provider deduplication unknown. Gonka token claims never become GPU stock or independently verified FLOP demand.
Compute-price forward signal
Kalshi monthly-average contracts · not rentable GPU stock
What it coversFive GPU price indices, not five physical GPUs. Each contract settles on one GPU model's monthly average USD/GPU-hour.
How to read odds“YES 72% at >$3” means the market price implies roughly a 72% chance of settling above $3. It is not a calibrated forecast.
How to use itCompare the median bracket with Ornn now. Higher suggests future scarcity; lower suggests waiting may pay. Use directionally, not as a rental quote.
GPU
Settlement month
Median price bracket
Ornn now
Implied move
Crossing odds
Volume
Open interest
View full contract ladders
GPU
Tenor
Contract
Best YES bid
Best YES ask
Last
Volume
Open interest
The median bracket spans adjacent strikes where YES midpoint odds cross 50%; its midpoint is compared with Ornn only for intuition, not presented as an expected price. Volume is contracts traded; open interest is outstanding contracts—neither measures GPUs or dollars. Spreads, fees, liquidity, and risk premia can distort probability-like prices.
Asynchronous batch offers
Separate quotes for 4,096 input + 512 output tokens, cold text, one request
Model / provider
Precision / context
Native token rates ($/1M)
Non-token rates
Workload quote
Evidence / constraints
Quotes preserve each endpoint's actual input, output, cache and non-token prices. Missing charges leave the total unknown. A comparison requires the same model, provider, precision, context and workload. No paid observations have been collected; billed charges are separate evidence.
Confirmed lower bound only: provider-reported units × Gn. Unknown stock is excluded, never assumed to be one or zero.
Confirmed capacity by provider
Current and time-weighted H100 equivalents
Exact provider-reported quantities only. “Quantity unknown” providers may have rentable inventory, but are excluded from H100e totals until their API exposes a defensible quantity or lower bound.
Capacity evidence
Counts and bounds for each SKU, region and topology
Provider / GPU
Native SKU
Region / zone
Topology
Interruptible
Availability
Method / confidence
Age
Bounds describe the observed scope and are never added to exact stock totals. Stale evidence displays unknown. Fleet bring-up, hardware validation, catalogs and account quotas prove no rentable units.
Price / Gn ranking
Qualified known capacity
Rank
Raw 7d
Rank = Σ(price × known units) ÷ Σ(known units) ÷ measured Gn. Raw floor is shown separately and never sets rank. Clore requires ≥95% reliability, ≥4.0 rating, and ≥5 ratings; unknown-unit offers are excluded. Host throttling remains unknown until measured; estimated Gn rows are unranked.
Provider network
Latest collection state
Providers needing credentials
Measured precision throughput
Dense GEMM throughput · TOPS
INT4
Measured matrix: 8192×8192 GEMM, 50 iterations. INT4 is not collected: the harness has no validated packed-INT4 kernel. NVFP4 is capability-gated and preserves unsupported/failed status. Gn ranking uses accepted BF16 throughput.
Full provider × GPU snapshot
Current status, price, units, H100e, Gn, and regions
Obtain the devnet operator's RPC/genesis, tailnet access, registered model hash and checkpoint revision, pinned worker/backend images, DA URL and funding requirements before renting.
On a compatible VM, check nvidia-smi, Docker GPU access, RAM and persistent disk. Use a dedicated miner signer stored privately on the host.
Measure the exact checkpoint and match its on-chain model root. Fill the private miner env template, then start the worker and one measured backend with the tracked Compose files.
Arrange the Ghost issuer entry and gateway allowlist. Wait for finalized calibration and capacity stake acceptance; check authenticated /session/info and a real accepted work request.
Keep the miner available through calibration and renewal. When retiring it, delete the rental in the provider console and verify billing stopped.
SOFT is the current development path. A healthy process or rented GPU does not establish readiness or earnings; HARD is not deployable from this repository.
What is compared, what is measured, and why API prices can beat public GPU rent
Planning model
Headline margin
The headline pools an equal one-million-token workload for every quoted model; it is not traffic-weighted because provider traffic shares are unavailable. A negative headline does not mean every model is negative.
OpenRouter uses the cheapest live endpoint with matching precision and at least 95% one-day uptime. Revenue blends standard input/output prices at the configured token mix. AIHubMix and SiliconFlow use their cheapest listed matching quote.
revenue = 80% input price + 20% output price
GPU cost
Each model gets its own capacity-weighted blend of qualified live rental stock. Eligible cards need a public price, measured Gn, compatible weights, and room for one full advertised-context KV/state cache.
Calibrated capacity is the measured BF16 dense-compute reference for one H100-equivalent hour. Benchmark-realized work converts the matched model's best measured input/output serving rates into Gn using its 2 × active-parameter sticker and the declared token mix.
gap = 1 − benchmark-realized Gₙ/hour ÷ calibrated Gₙ/hour
Example: an 84.27% gap means the benchmark delivered 15.73% of calibrated capacity as model-token work. The remainder reflects model/runtime delivery limits such as memory movement, attention, KV-cache traffic, kernel shapes, and serial decode. It is not 84.27% GPU idle time and does not measure production fleet utilization or idle lease time.
Prompt caching
Not included in the headline. The table compares published route-level input, cache-read, and cache-write prices. Missing cache prices stay unavailable, never zero. Hit-rate rows blend cold input and steady-state reads; they exclude initial writes, TTL/eviction, and reuse distribution because providers do not publish a traffic-wide hit share. Warm self-host cost appears only after the exact model, quantization, GPU, full context, repeated prefix, and vLLM cache counters are measured.
Our cost basis is public on-demand rental economics. API providers may own or reserve hardware, use TPUs or undisclosed GPUs, batch more requests, sustain higher utilization, reuse prefixes, or price below public rental cost. This is not asserted to be like-for-like hardware.
Evidence and uncertainty
Loading benchmark coverage…
Full-context session counts are memory ceilings, not measured concurrency. Sensitivity ranges vary efficiency, utilization, and overhead; they are not confidence intervals. AIHubMix units remain unvalidated.